Improve M2 ref conf compute performance

[ldgauthier]

The somatic likelihoods model becomes very expensive when evaluated at every reference position. MT calling takes about 40 minutes in GVCF mode and 5 minutes without ref conf. Ideas for optimization include:

• Reducing the convergence threshold for alleleFractionsPosterior method (my initial attempts here didn't make a big improvement)
• Using a simplified likelihoods model for reference sites (similar to lnLikelihoodRatio used for active region determination)
• Keeping LOD calculation from active region determination and only calculating it once